Allocating Memory
Ted Baker  Andy Wang
CIS 4930 / COP 5641
Topics
 kmalloc and friends
 get_free_page and “friends”
 vmalloc and “friends”
 Memory usage pitfalls
Linux Memory Manager (1)
 Page allocator maintains individual pages
Page allocator
Linux Memory Manager (2)
 Zone allocator allocates memory in power-of-
two sizes
Zone allocator
Page allocator
Linux Memory Manager (3)
 Slab allocator groups allocations by sizes to
reduce internal memory fragmentation
Slab allocator
Zone allocator
Page allocator
kmalloc
 Does not clear the memory
 Allocates consecutive virtual/physical memory
pages

Offset by PAGE_OFFSET

No changes to page tables
 Tries its best to fulfill allocation requests

Large memory allocations can degrade the
system performance significantly
The Flags Argument
 kmalloc prototype
#include <linux/slab_def.h>
void *kmalloc(size_t size, int flags);
 GFP_KERNEL is the most commonly used flag



Eventually calls __get_free_pages (the origin of
the GFP prefix)
Can put the current process to sleep while waiting for a
page in low-memory situations
Cannot be used in atomic context
The Flags Argument

To obtain more memory
 Flush dirty buffers to disk
 Swapping out memory from user processes
 GFP_ATOMIC is called in atomic context



Interrupt handlers, tasklets, and kernel timers
Does not sleep
If the memory is used up, the allocation fails
 No flushing and swapping
 Other flags are available
 Defined in <linux/gfp.h>
The Flags Argument
 GFP_USER is used to allocate user pages; it may
sleep
 GFP_HIGHUSER allocates high memory user pages
 GFP_NOIO disallows I/O
 GFP_NOFS does not allow making file system calls


Used in file system and virtual memory code
Disallow kmalloc to make recursive calls to file
system and virtual memory code
The Flags Argument
 Allocation priority flags
 Prefixed with __
 Used in combination with GFP flags (via ORs)
 __GFP_DMA requests allocation to happen in the
DMA-capable memory zone
 __GFP_HIGHMEM indicates that the allocation may be
allocated in high memory
 __GFP_COLD requests for a page not used for some
time (to avoid DMA contention)
 __GFP_NOWARN disables printk warnings when an
allocation cannot be satisfied
The Flags Argument
 __GFP_HIGH marks a high priority request
 Not for kmalloc
 __GFP_REPEAT

Try harder
 __GFP_NOFAIL

Failure is not an option (strongly discouraged)
 __GFP_NORETRY

Give up immediately if the requested memory is not
available
Memory Zones
 DMA-capable memory



Platform dependent
First 16MB of RAM on the x86 for ISA devices
PCI devices have no such limit
 Normal memory
 High memory


Platform dependent
> 32-bit addressable range
Memory Zones
 If __GFP_DMA is specified

Allocation will only search for the DMA zone
 If nothing is specified

Allocation will search both normal and DMA
zones
 If __GFP_HIGHMEM is specified

Allocation will search all three zones
The Size Argument
 Kernel manages physical memory in pages

Needs special management to allocate small
memory chunks
 Linux creates pools of memory objects in
predefined fixed sizes (32-byte, 64-byte, 128byte memory objects)
 Smallest allocation unit for kmalloc is 32 or
64 bytes
 Largest portable allocation unit is 128KB
Lookaside Caches (Slab Allocator)
 Nothing to do with TLB or hardware caching
 Useful for USB and SCSI drivers

Improved performance
 To create a cache for a tailored size
#include <linux/slab.h>
kmem_cache_t *
kmem_cache_create(const char *name, size_t size,
size_t offset, unsigned long flags,
void (*constructor) (void *, kmem_cache_t *,
unsigned long flags),
void (*destructor) (void *, kmem_cache_t *,
unsigned long flags));
Lookaside Caches (Slab Allocator)
 name: memory cache identifier

Allocated string without blanks
 size: allocation unit
 offset: starting offset in a page to align
memory

Most likely 0
Lookaside Caches (Slab Allocator)
 flags: control how the allocation is done
 SLAB_NO_REAP



Prevents the system from reducing this memory
cache (normally a bad idea)
Obsolete
SLAB_HWCACHE_ALIGN



Requires each data object to be aligned to a
cache line
Good option for frequently accessed objects on
SMP machines
Potential fragmentation problems
Lookaside Caches (Slab Allocator)

SLAB_CACHE_DMA


Requires each object to be allocated in the DMA
zone
See mm/slab.h for other flags
 constructor: initialize newly allocated
objects
 destructor: clean up objects before an
object is released
 Constructor/destructor may not sleep due to
atomic context
Lookaside Caches (Slab Allocator)
 To allocate an memory object from the
memory cache, call
void *kmem_cache_alloc(kmem_cache_t *cache, int flags);

cache: the cache created previously
flags: same flags for kmalloc

Failure rate is rather high


Must check the return value
 To free an memory object, call
void kmem_cache_free(kmem_cache_t *cache,
const void *obj);
Lookaside Caches (Slab Allocator)
 To free a memory cache, call
int kmem_cache_destroy(kmem_cache_t *cache);


Need to check the return value
Failure indicates memory leak
 Slab statistics are kept in /proc/slabinfo
A scull Based on the Slab Caches:
scullc
 Declare slab cache
kmem_cache_t *scullc_cache;
 Create a slab cache in the init function
/* no constructor/destructor */
scullc_cache
= kmem_cache_create("scullc", scullc_quantum, 0,
SLAB_HWCACHE_ALIGN, NULL, NULL);
if (!scullc_cache) {
scullc_cleanup();
return -ENOMEM;
}
A scull Based on the Slab Caches:
scullc
 To allocate memory quanta
if (!dptr->data[s_pos]) {
dptr->data[s_pos] = kmem_cache_alloc(scullc_cache,
GFP_KERNEL);
if (!dptr->data[s_pos])
goto nomem;
memset(dptr->data[s_pos], 0, scullc_quantum);
}
 To release memory
for (i = 0; i < qset; i++) {
if (dptr->data[i]) {
kmem_cache_free(scullc_cache, dptr->data[i]);
}
}
A scull Based on the Slab Caches:
scullc
 To destroy the memory cache at module
unload time
/* scullc_cleanup: release the cache of our quanta */
if (scullc_cache) {
kmem_cache_destroy(scullc_cache);
}
Memory Pools
 Similar to memory cache


Reserve a pool of memory to guarantee the
success of memory allocations
Can be wasteful
 To create a memory pool, call
#include <linux/mempool.h>
mempool_t *mempool_create(int min_nr,
mempool_alloc_t *alloc_fn,
mempool_free_t *free_fn,
void *pool_data);
Memory Pools
 min_nr is the minimum number of allocation
objects
 alloc_fn and free_fn are the allocation
and freeing functions
typedef void *(mempool_alloc_t)(int gfp_mask,
void *pool_data);
typedef void (mempool_free_t)(void *element,
void *pool_data);
 pool_data is passed to the allocation and
freeing functions
Memory Pools
 To allow the slab allocator to handle
allocation and deallocation, use predefined
functions
cache = kmem_cache_create(...);
pool = mempool_create(MY_POOL_MINIMUM, mempool_alloc_slab,
mempool_free_slab, cache);
 To allocate and deallocate a memory pool
object, call
void *mempool_alloc(mempool_t *pool, int gfp_mask);
void mempool_free(void *element, mempool_t *pool);
Memory Pools
 To resize the memory pool, call
int mempool_resize(mempool_t *pool, int new_min_nr,
int gfp_mask);
 To deallocate the memory poll, call
void mempool_destroy(mempool_t *pool);
get_free_page and Friends
 For allocating big chunks of memory, it is
more efficient to use a page-oriented allocator
 To allocate pages, call
/* returns a pointer to a zeroed page */
get_zeroed_page(unsigned int flags);
/* does not clear the page */
__get_free_page(unsigned int flags);
/* allocates multiple physically contiguous pages */
__get_free_pages(unsigned int flags, unsigned int order);
get_free_page and Friends
 flags

Same as flags for kmalloc
 order

Allocate 2order pages


order = 0 for 1 page
order = 3 for 8 pages

Can use get_order(size)to find out order

Maximum allowed value is about 10 or 11
See /proc/buddyinfo statistics

get_free_page and Friends
 Subject to the same rules as kmalloc
 To free pages, call
void free_page(unsigned long addr);
void free_pages(unsigned long addr, unsigned long order);
 Make sure to free the same number of pages

Or the memory map becomes corrupted
A scull Using Whole Pages:
scullp
 Memory allocation
if (!dptr->data[s_pos]) {
dptr->data[s_pos] =
(void *) __get_free_pages(GFP_KERNEL, dptr->order);
if (!dptr->data[s_pos])
goto nomem;
memset(dptr->data[s_pos], 0, PAGE_SIZE << dptr->order);
}
 Memory deallocation
for (i = 0; i < qset; i++) {
if (dptr->data[i]) {
free_pages((unsigned long) (dptr->data[i]), dptr->order);
}
}
The alloc_pages Interface
 Core Linux page allocator function
struct page *alloc_pages_node(int nid, unsigned int flags,
unsigned int order);

nid: NUMA node ID
 Two higher level macros
struct page *alloc_pages(unsigned int flags,
unsigned int order);
struct page *alloc_page(unsigned int flags);

Allocate memory on the current NUMA node
The alloc_pages Interface
 To release pages, call
void __free_page(struct page *page);
void __free_pages(struct page *page, unsigned int order);
/* optimized calls for cache-resident or non-cache-resident
pages */
void free_hot_page(struct page *page);
void free_cold_page(struct page *page);
vmalloc and Friends
 Allocates a virtually contiguous memory
region




Not consecutive pages in physical memory

Each page retrieved with a separate
alloc_page call

Less efficient
Can sleep (cannot be used in atomic context)
Returns 0 on error, or a pointer to the
allocated memory
Its use is discouraged
vmalloc and Friends
 vmalloc-related prototypes
#include <linux/vmalloc.h>
void
void
void
void
*vmalloc(unsigned long size);
vfree(void * addr);
*ioremap(unsigned long offset, unsigned long size);
iounmap(void * addr);
vmalloc and Friends
 Each allocation via vmalloc involves setting
up and modifying page tables

Return address range between
VMALLOC_START and VMALLOC_END
(defined in <linux/pgtable.h>)

Used for allocating memory for a large
sequential buffer
vmalloc and Friends
 ioremap builds page tables


Does not allocate memory
Takes a physical address (offset) and return
a virtual address


Useful to map the address of a PCI buffer to
kernel space
Should use readb and other functions to
access remapped memory
A scull Using Virtual Addresses:
scullv
 This module allocates 16 pages at a time
 To obtain new memory
if (!dptr->data[s_pos]) {
dptr->data[s_pos]
= (void *) vmalloc(PAGE_SIZE << dptr->order);
if (!dptr->data[s_pos]) {
goto nomem;
}
memset(dptr->data[s_pos], 0,
PAGE_SIZE << dptr->order);
}
A scull Using Virtual Addresses:
scullv
 To release memory
for (i = 0; i < qset; i++) {
if (dptr->data[i]) {
vfree(dptr->data[i]);
}
}
Per-CPU Variables
 Each CPU gets its own copy of a variable


Almost no locking for each CPU to work with
its own copy
Better performance for frequent updates
 Example: networking subsystem


Each CPU counts the number of processed
packets by type
When user space request to see the value,
just add up each CPU’s version and return the
total
Per-CPU Variables
 To create a per-CPU variable
#include <linux/percpu.h>
DEFINE_PER_CPU(type, name);
 name: an array
DEFINE_PER_CPU(int[3], my_percpu_array);

Declares a per-CPU array of three integers
 To access a per-CPU variable

Need to prevent process migration
get_cpu_var(name); /* disables preemption */
put_cpu_var(name); /* enables preemption */
Per-CPU Variables
 To access another CPU’s copy of the
variable, call
per_cpu(name, int cpu_id);
 To dynamically allocate and release per-CPU
variables, call
void *alloc_percpu(type);
void *__alloc_percpu(size_t size);
void free_percpu(const void *data);
Per-CPU Variables
 To access dynamically allocated per-CPU
variables, call
per_cpu_ptr(void *per_cpu_var, int cpu_id);
 To ensure that a process cannot be moved
out of a processor, call get_cpu (returns cpu
ID) to block preemption
int cpu;
cpu = get_cpu()
ptr = per_cpu_ptr(per_cpu_var, cpu);
/* work with ptr */
put_cpu();
Per-CPU Variables
 To export per-CPU variables, call
EXPORT_PER_CPU_SYMBOL(per_cpu_var);
EXPORT_PER_CPU_SYMBOL_GPL(per_cpu_var);
 To access an exported variable, call
/* instead of DEFINE_PER_CPU() */
DECLARE_PER_CPU(type, name);
 More examples in
<linux/percpu_counter.h>
Obtaining Large Buffers
 First, consider the alternatives


Optimize the data representation
Export the feature to the user space
 Use scatter-gather mappings
 Allocate at boot time
Acquiring a Dedicated Buffer at Boot
Time
 Advantages


Least prone to failure
Bypass all memory management policies
 Disadvantages



Inelegant and inflexible
Not a feasible option for the average user
Available only for code linked to the kernel

Need to rebuild and reboot the computer to install
or replace a device driver
Acquiring a Dedicated Buffer at Boot
Time
 To allocate, call one of these functions
#include <linux/bootmem.h>
void *alloc_bootmem(unsigned long size);
/* need low memory for DMA */
void *alloc_bootmem_low(unsigned long size);
/* allocated in whole pages */
void *alloc_bootmem_pages(unsigned long size);
void *alloc_bootmem_low_pages(unsigned long size);
Acquiring a Dedicated Buffer at Boot
Time
 To free, call
void free_bootmem(unsigned long addr, unsigned long size);
 Need to link your driver into the kernel
 See Documentation/kbuild
Memory Usage Pitfalls
 Failure to handle failed memory allocation

Needed for every allocation
 Allocate too much memory
 No built-in limit on memory usage